Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Cloudflare Releases Clairic for Fast Value Extraction - OpenSmartRoute
Cloudflare Releases Clairic for Fast Value Extraction
Cloudflare launched Clairic, a model that extracts typed values like numbers and dates in under half a second. It runs on Workers AI and costs less than half as much as previous all-LLM pipelines for similar tasks.
Key points
Extraction took 0.46 seconds versus 16.1 seconds for GLM-5.3.
Cost per million claims dropped 49% to $6,032.
The model handles up to four images per request.
Every answer includes a confidence score or null probability.
Why it matters: Teams save money and speed by replacing slow reasoning models with fast extraction tools for invoices and claims.
By OpenSmartRoute editorial · written through the router by writer-small
From MarkTechPost - “Cloudflare Releases Clairic, a System One Extraction Model on Workers AI”
Cloudflare released Clairic as a preview model today. This new system runs on Workers AI. It extracts typed values from text or images. The model returns numbers, dates, and IDs directly. It does not generate natural language text. Developers can use it to pull data from invoices or emails.
Clairic answers every requested field in one forward pass. It uses no reasoning steps during this process. The output contains no generated tokens. This makes the operation extremely fast. The system reads the input and maps it to fields. Each field has a specific type like number or date.
The model supports up to four images per request. It handles a context window of 65,536 tokens. Parameter counts remain undisclosed by Cloudflare. Users must run it on Cloudflare-hosted GPUs. Local deployment is not possible until weights ship. The official binding is @cf/cloudflare/clairic.
Clairic targets a specific job often done by LLMs. It pulls totals, dates, and IDs into JSON format. Unlike general models, it does not write paragraphs. It returns structured data for automation. This fits well with agents that need precise inputs.
The system builds on the System One idea. This concept comes from TypeSafe's Jev project. It is named after Kahneman's fast, intuitive thinking. Clairic handles answers you cannot list in advance. Examples include invoice totals or repair costs. It is distinct from decision models like Clef.
Cloudflare frames a three-part stack for complex tasks. Extraction models turn raw input into typed values. Decision models turn those values into choices. Large language models turn decisions into words. Clairic sits at the first step of this chain.
A request has a state containing text and images. It also has a map of named, typed fields. Each field can carry plain-language instructions. Supported types include text, number, date, and list. The Workers AI changelog lists these specific types.
Every scalar answer includes a probability for absence. This is called p_null. It also includes a confidence score. Text answers return exact spans with character offsets. Estimates and forecasts return a median and an 80% interval. Forecasts cover up to 64 steps.
There is no field limit on the model. Fields that do not fit in one pass split into parallel passes. Each pass sees the whole input data. In a docs example, a 61-token invoice ran in three passes. It billed 655 input tokens total. This included 472 schema tokens.
Deepseek released V4.1 Flash 1017 which beats larger rivals in coding and agent tasks. The model also cuts per-task costs significantly while producing shorter outputs.
doubleAI raised $200 million to sell an AI model that optimizes software and hardware. The model speeds up code and lowers hardware usage.
Wide schemas can cost more than the document itself. The input tokens count includes schema tokens. Developers must watch their token usage carefully. The billing model charges per million input tokens. Output tokens cost nothing for this model.
In a support-email example, Clairic added price and tax. It computed 11 days since delivery. Neither number appeared in the original email. A missing tracking number returned null instead. The model did not guess a plausible value.
On a skewed receipt photo, it read seven present values. It returned null for the absent discount. Its lowest confidence was 0.776 on the order number. This shows the model handles missing data gracefully. It does not hallucinate numbers when unsure.
For forecasting, Cloudflare predicted daily TSA checkpoint counts. It covered 21 days of data. Average error was 3.3% across the forecast. 17 of 21 actual counts fell inside the 80% interval. 3 of the 4 misses fell on Columbus Day weekend.
The main test was a damaged-product claims pipeline. Clairic extracted facts, Clef chose refund or repair. GLM-5.3 wrote the reply text. Cloudflare compared this against GLM-5.3 handling all three steps. The cascade finished in 19.3 seconds.
The previous pipeline took 61.7 seconds. That was about 3.2 times slower. Extraction took 0.46 seconds against 16.1 seconds. The decision step took 0.22 seconds against 21.9 seconds. Cost per million claims fell 49%.
Cost dropped from $11,740 to $6,032. The reply step still dominated the time. It took 18.6 of the 19.3 seconds. Both pipelines got the facts and decision right. Clairic's repair estimate was identical across runs.
GLM-5.3 returned $250, $250, and $220 in tests. Cloudflare says Gemma 4 and Qwen failed the all-LLM version. Clairic provides consistent results for extraction tasks. It avoids the variability of generation models.
Typed values include spans, numbers, dates, and durations. Typed decisions have calibrated probabilities. The extraction step took 0.46 seconds in a Cloudflare test. The previous step took 16.1 seconds for comparison. Clairic is available as a Workers AI preview. Open weights are planned for release.
Developers can use Clairic through the Workers AI binding. They call it with env.AI.run() or the REST API. AI Gateway works with the same endpoints. The model is live for testing now.
Pricing is $0.10 per million input tokens. There is no output charge. This undercuts Clef's $0.24 price. It sits above Jev's $0.042 rate. Weights are promised but not yet released.
Clairic returns typed values, not text. It does this in one forward pass. This speed is significant for high-volume tasks. The cost structure is very favorable for extraction.
Every answer carries confidence, p_null, spans, or intervals. Users can check the Cloudflare blog post for details. The Clairic model docs explain the API. The extraction models explainer covers the technology. The Workers AI changelog lists the types.
Developers should read the earlier coverage of Drex 1.5. All credit goes to the researcher of this project. You can follow the team on Twitter. Join the 150k+ ML SubReddit community. Subscribe to the newsletter for updates.
Check the MarkTechPost article for the full story. Asif Razzaq is the CEO of Marktechpost. He is an engineer and entrepreneur. His platform covers machine learning news. It has over 2 million monthly views.
The article provides deep technical details. It explains how the model works internally. You can try the preview model now. Test it with your own invoice data. Compare the speed against your current pipeline.
Measure the cost savings in your own environment. Track the tokens used for schema vs input. See how the confidence scores look for your data.
OpenSmartRoute helps teams route requests efficiently. It sends each request to the best-fit model. The router scores candidates on quality, cost, speed, and safety. Teams set the weights per request.
Hard rules pin a request to specific conditions. Text with personal data stays on an on-premises model. A region or cost cap is never crossed. The system learns from outcomes automatically.
A model that answers well gets more traffic. One that fails gets less traffic. A new model is one catalogue entry. It competes on the next request. Nothing else changes in the app.
The router works with any OpenAI-compatible provider. It supports open-weight models served locally. It handles MCP tools and A2A agents. A savings ledger shows what each routed request cost.
The ledger compares the cost to the most expensive model. An input guard spots prompt injection before a request leaves. The 'osr eval' tool measures routing accuracy. It can fail a build when it drops.
The hosted platform keeps a models catalogue. It has prices and public rankings built from real traffic. Teams gain immediate access to Clairic through the router. They avoid managing the model directly.
The router handles the complexity of model selection. Developers focus on their application logic. Clairic becomes a standard tool in the stack. The team sees real-time cost savings.
Testing Clairic now is the best next step. Integrate it into your extraction pipeline. Monitor the confidence scores for your data. Watch the cost reduction over time.
The preview allows safe experimentation without risk. You can test with small datasets first. Scale up once you are confident in the results.
Cloudflare's announcement marks a shift in model design. Extraction is now a specialized, fast model. It replaces the need for general LLMs for this task.
This change improves speed and reduces costs significantly. Agents can rely on Clairic for data extraction. The system is ready for production use soon.
Teams should evaluate their current extraction workflows. Look for opportunities to use Clairic. The savings could be substantial for high-volume tasks.
The model handles complex inputs like damaged receipts. It manages missing data without hallucinating. This reliability is crucial for automated systems.
Forecasting capabilities add another layer of utility. The 80% interval gives a clear range of uncertainty. Users can plan based on these predictions.
The three-part stack offers a complete solution. Extraction, decision, and generation are now distinct. This modularity improves overall system performance.
Clairic is a powerful addition to the AI toolkit. It solves a specific problem very well. Developers should consider adopting it for their projects.
The preview phase is the right time to try it. Gather feedback and refine your use cases. The open weights will be available soon.
Keep an eye on the Workers AI changelog. New features might arrive quickly. The community will likely share more insights.
This news highlights the trend toward specialized models. General LLMs are becoming less efficient for simple tasks. Clairic proves that speed and cost matter.
Engineers can build faster agents with this tool. Managers can see clear cost reductions in their ledgers. Both groups benefit from this technology.
The integration with OpenSmartRoute makes adoption easier. The router handles the routing logic automatically. Teams save time on infrastructure management.
Final steps involve testing and deployment. Start with a small batch of data. Measure the improvements in speed and cost.
The future of AI extraction looks bright. Specialized models will replace general ones for specific jobs. Clairic is a step in that direction.
Stay updated with Cloudflare's releases. Follow the progress of the open weights. Join the community discussions on the topic.
The technology is ready for real-world use. Just plug it into your pipeline and test. The results should show significant improvements.
Cloudflare has set a new standard for extraction. Other companies will likely follow this path. The industry is moving toward specialization.
This release is a major milestone for the sector. It shows the power of system one thinking. Clairic is a model to watch closely.
Developers should explore the API documentation. Understand the schema and token costs. Plan your architecture around these constraints.
The model handles images and text well. It is robust for unstructured inputs. This makes it versatile for many industries.
The cost structure is very competitive. It beats previous all-LLM pipelines significantly. This is a key selling point.
Speed is another major advantage. Sub-second extraction is rare for this task. It enables real-time processing capabilities.
The confidence scores help users trust the output. They know when the model is unsure. This transparency is valuable for safety.
The forecast interval provides statistical rigor. Users can make informed decisions based on data. It reduces the risk of bad guesses.
Clairic fits into the Cloudflare ecosystem perfectly. It runs on the same infrastructure as other tools. This simplifies deployment for customers.
The preview model is limited to Cloudflare-hosted GPUs. Local options will come later. Plan your infrastructure accordingly for now.
The team behind Clairic is doing great work. They solved a hard problem elegantly. Their approach is worth studying.
The article details the performance numbers clearly. You can see the exact speed improvements. The cost savings are also quantified.
This is a practical tool for developers. It solves a real business problem. Adoption should be high once tested.
The integration with Workers AI is seamless. No extra infrastructure is needed. This lowers the barrier to entry.
The model supports many data types. It handles lists, estimates, and forecasts. This flexibility is a big plus.
The p_null probability is a unique feature. It tells you if a value is missing. This is useful for data cleaning.
The confidence score adds another layer of insight. Users can filter results based on reliability. This improves the quality of downstream tasks.
The parallel passes for wide schemas are smart. They avoid losing data in long inputs. This ensures accuracy for complex documents.
The comparison with GLM-5.3 shows the gap. The speed difference is dramatic. The cost difference is also huge.
The decision step improvement is notable too. It took 0.22 seconds against 21.9 seconds. The whole pipeline is much faster.
The consistency of results is impressive. Clairic gave the same estimate across runs. This reliability is hard to find in LLMs.
The failure of Gemma 4 and Qwen is a warning. General models are not always reliable for extraction. Specialized models are the better choice.
The Cloudflare team did a great job here. They built a tool that developers need. The market will likely embrace it quickly.
The preview allows for early adoption. Teams can get ahead of the curve. They can shape the future of the tool.
The open weights promise is exciting. It will make the model accessible to everyone. This will drive further innovation.
The community will likely build many use cases. From finance to logistics, the applications are endless. Clairic is a foundational model for many.
The story of Clairic is a story of efficiency. It proves that specialized models win. The future is here and it is fast.
Engineers should start planning their migration. Managers should approve the budget for it. The ROI is clear and strong.
The OpenSmartRoute router makes this migration easy. It handles the routing and cost tracking. Teams can focus on their core logic.
The savings ledger will show the value immediately. You will see the cost drop in real time. This data is powerful for decision making.
The input guard protects your data from leaks. This is a critical feature for security. It ensures compliance with regulations.
The 'osr eval' tool ensures the router works well. It tests the routing accuracy on your prompts. This prevents bad routing in production.
The hosted platform provides a reliable backend. You do not need to manage the servers. This reduces operational overhead significantly.
The catalogue of models grows with every release. Clairic is a new and powerful entry. It competes on quality and cost.
The team at Cloudflare is leading the way. They are setting the pace for the industry. Others will follow their lead.
This release is a must-read for everyone. It changes how we think about extraction. It is a game changer for efficiency.
The technology is mature and ready. Just integrate it and start using it. The benefits will be immediate and clear.
Cloudflare has delivered on its promise. Clairic is a fast, cheap, and reliable tool. It is exactly what the market needs.
The preview phase is the perfect time to try. Test it thoroughly before the full release. Gather your data and run the tests.
The model handles edge cases well. It returns null for missing data. It does not make things up.
The forecast accuracy is solid for TSA counts. The 3.3% average error is very low. The 80% interval is reliable.
The integration with Clef and GLM-5.3 is smooth. The three-step pipeline works together perfectly. The end result is a fast, cheap agent.
The story of Clairic is a testament to engineering. It shows the power of focused model design. It is a model to emulate.
The future of AI is specialized and efficient. Clairic is a step in that direction. It is a model to watch.
Engineers and managers should pay attention. This news affects their workflows and budgets. It is important to understand it.
The OpenSmartRoute router is the perfect companion. It maximizes the value of every model. It ensures you get the best performance.
The team can set up the router in minutes. The catalogue is ready for Clairic. The savings ledger will track the gains.
The input guard protects against prompt injection. This is a vital security feature. It keeps your data safe from attacks.
The 'osr eval' tool ensures quality control. It fails the build if routing drops. This prevents bad deployments.
The hosted platform is easy to use. The models catalogue is well maintained. The prices are transparent and accurate.
The router learns from outcomes automatically. It optimizes the traffic distribution over time. This makes the system smarter every day.
The team can define the rules for their app. They can set cost caps and safety rules. This gives them full control.
The integration with OpenAI-compatible providers is flexible. You can mix and match models easily. This supports diverse use cases.
The open-weight models can be served locally. This gives teams full data sovereignty. It meets strict compliance requirements.
The MCP tools and A2A agents are supported. The router is ready for the future. It scales with your needs.
The savings ledger is a powerful tool. It shows the cost of every request. It helps you optimize your spending.
The input guard is a simple but effective feature. It stops bad data from entering the system. It is a must-have for production.
The 'osr eval' tool is essential for quality. It measures the routing accuracy on your prompts. It ensures the system works as expected.
The hosted platform keeps the models catalogue updated. It reflects the real traffic and prices. This ensures you are always getting the best deal.
The team at OpenSmartRoute is committed to quality. The router is built for reliability and performance. It is a tool for serious engineers.
The integration with Clairic is seamless. The router handles the routing logic automatically. It saves the team time and effort.
How OpenSmartRoute helps
Teams using OpenSmartRoute gain immediate savings by routing Clairic requests to the new model instead of expensive general LLMs. The router scores candidates on cost, speed, and quality before sending each request. This ensures the team pays only $6,032 per million claims rather than higher rates from older pipelines.
The setup requires adding Clairic to the existing models catalogue with its specific price and speed metrics. Hard rules can enforce that text containing personal data stays on secure on-premises models while Clairic handles public data. The input guard automatically spots prompt injection attempts before any request leaves the system.
A savings ledger will show the cost difference next to what the most expensive model would have charged. The 'osr eval' tool measures routing accuracy on the team's own prompts to ensure the new model performs well. Teams can update the catalogue entry and let the router compete Clairic on the next request automatically.
Mistral AI launched Timestral, a new family of time series forecasting models ranging from 5M to 209M parameters. The largest model outperforms Google's TimesFM-3 on key benchmarks while running on standard server CPUs.
